Research map / Large language models / Retrieval and knowledge

KG-Guard: Graph-Based Hallucination Detection for Knowledge Base Question Answering

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Knowledge graph reasoning with LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes KG-Guard, a lightweight black-box graph-encoder detector that classifies LLM-proposed answer nodes in KBQA subgraphs as hallucinated or factual, and shows its flags enable iterative answer refinement.

Problem

Detecting hallucinated answer nodes returned by LLM-based knowledge base question answering over retrieved KG subgraphs, without access to LLM internals.

Approach

Formulates KBQA hallucination detection as per-answer-node classification on an augmented question-specific subgraph: semantic node features plus learned topic/answer marks, a virtual question node linked to topic entities, a graph transformer encoder, and an MLP head with the question embedding.

Key results

• KG-Guard (GraphTransformer) achieves F1 82.0 on WebQSP, 87.4 on CWQ, and 84.3 on PUGG, the highest among compared methods.
• Iterative refinement guided by KG-Guard flags raises KBQA F1 by 13.0-14.5 points and Exact Match by 16.9-17.6 points on the three datasets.
• KG-Guard adds 2.52M trainable parameters on a frozen ~355M encoder, versus Llama-4-Scout at 109B total parameters, with zero LLM calls at detection time.

Surprising findings

• All three Llama-4-Scout judge variants score below the Random baseline on WebQSP (F1 36.5-41.0 vs 49.6).
• Training on hallucination labels outperforms training on KBQA gold labels with identical training set sizes.

Limitations

• Cannot handle LLM abstentions because they are not represented as graph nodes.
• Evaluated on a single KBQA pipeline (KAPING with Llama-4-Scout), so generalization to other pipelines is untested.
• Requires training examples with hallucination labels.

Experimental setup

Datasets: WebQSP ComplexWebQuestions PUGG

Benchmarks: WebQSP ComplexWebQuestions PUGG

Metrics: F1 Precision Recall Accuracy AUC-PR Exact Match

Baselines: GPT-5-mini (LLM-as-judge) Llama-4-Scout (LLM-as-judge) SelfCheckGPT Most Frequent Random GAT variant of KG-Guard

Builds on: KAPING G-Retriever GraphTransformer SelfCheckGPT GAT

Closest papers by meaning

• Debate on Graph: A Flexible and Reliable Reasoning Framework for Large Language Models (AAAI 2025 · similarity 0.91)
• Diagnosing and Addressing Pitfalls in KG-RAG Datasets: Toward More Reliable Benchmarking (NeurIPS 2025 · similarity 0.90)
• Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation (AAAI 2025 · similarity 0.89)
• KBQA-o1: Agentic Knowledge Base Question Answering with Monte Carlo Tree Search (ICML 2025 · similarity 0.88)
• FlexKBQA: A Flexible LLM-Powered Framework for Few-Shot Knowledge Base Question Answering (AAAI 2024 · similarity 0.88)
• Knowledge Editing with Dynamic Knowledge Graphs for Multi-Hop Question Answering (AAAI 2025 · similarity 0.88)